---
title: 'In the News: October 28, 2026 (Extra 3)'
description: "Funnel's A/B test: GLM made 66% more tool calls than Opus and finished 14% slower. An interview study of two dozen engineers finds stock harnesses dominate."
canonical_url: 'https://darkfactory.dev/news/2026-10-28-extra-3'
markdown_url: 'https://darkfactory.dev/news/2026-10-28-extra-3.md'
collection: news
date_published: '2026-10-28T23:05:00-04:00'
date_modified: '2026-10-28T23:05:00-04:00'
---

# In the News: October 28, 2026 (Extra 3)


Funnel ran a randomised test of an open-weight model against Claude Opus in its own data agent. The faster model made more tool calls and finished later, and one run repeated a single successful query 254 times. A separate write-up of interviews with about two dozen engineers reports that almost none of them run swarms or software factories.

## 1. Funnel: the model with faster calls finished its tasks 14% later

**[Our AI Agent Executed the Same Successful SQL Query 254 Times](https://armincat.com/articles/ai-agent-executed-same-query-254-times)** · Funnel (page shows no byline) · armincat.com, October 28, 2026

Funnel randomly assigned each new task in Reef, its internal data agent, to either GLM-5.2 (open weights) or Claude Opus 4.8. GLM averaged 7.8 seconds per model call against 18.1 for Opus. End to end it made 66% more tool calls (19 against 11 on average) and took 14% longer to finish. In two runs it executed the same successful SQL query 113 and 254 times. In one case the tool exposed only a 20-row preview, and the model kept re-running the query instead of noticing the limit. Funnel says it saw similar behavior from Kimi K2.6 and Qwen3.8 27B, with Qwen about twice as aggressive in tool use as GLM.

Funnel does not put the blame on the model alone. Some saved agent instructions named Athena tables and fields that did not exist, tool contracts were ambiguous, and nothing stopped repeated calls. In its words, "A successful tool call is not necessarily a useful tool call." The post gives no task count or variance, so the figures are one company's averages, and Funnel says it now runs GLM-5.3.

**Why it matters:** Per-call latency would have ranked these two models the wrong way round. If a team swaps models in a harness, the checks that matter are repeated-call detection, an execution budget and tool results that say whether they are complete, because a weaker model will find the gaps a stronger one papered over.

## 2. Ilya Razenshteyn: interviews with about two dozen engineers find stock tools and human review

**[Pragmatic guide to agentic coding](https://ilyaraz.substack.com/p/pragmatic-guide-to-agentic-coding)** · Ilya Razenshteyn, formerly of OpenAI, Pyte and Microsoft Research · Substack, October 26, 2026

Razenshteyn interviewed what he calls a couple of dozen engineers and researchers, from frontier labs to finance to small start-ups, about how they use coding agents at work. He dropped a planned survey because the stories were too varied, and he calls the result "qualitative and somewhat vibes-based." He does not say how interviewees were chosen.

He reports that almost everyone uses Claude or GPT models, and that harnesses are "almost exclusively" stock Claude Code or Codex. No one described swarms of agents, software factories or Gas Town-style setups; he gives cost and a lack of payoff as the reasons. The usual loop is plan, review the plan, implement, review the pull request with an agent and sometimes by hand, then merge. Several interviewees use one model to review another's work.

On correctness, he says code can go unread only for demos and one-off scripts, for strictly testable code such as a CUDA kernel, and for research code that only needs to be directionally correct. Elsewhere interviewees still read code or tests, and almost universally have some human peer review. The two failure modes he lists are messy code that accumulates as debt, and an agent misreading a requirement while the tests it writes reinforce the misreading. He also reports "slop" pull requests and design documents, and internal tools built in a day and then abandoned, as the main team-level problems.

**Why it matters:** This is self-reported and unmeasured, so it does not show what most teams do. It does give a dated counterpoint to factory-style setups from people working inside companies, and its practical rule is one a team can adopt now: mark which parts of a repository are reviewed and which are not.

---

## Also this cycle

- **[Hermit](https://github.com/Infini-AI-Lab/hermit)** · Infini-AI-Lab · An Apache 2.0 framework for training a "harness adapter", an external layer that adapts an existing agent without changing its model or harness. It lists GEPA, SkillOpt, MetaWorld Model, Harness² and Continual Harness as supported methods. The README's own demo says its score gain is constructed and "not a real-model performance result"; the per-benchmark results in the repository were not read. It had 292 stars when read.